Genome-Wide Functional Annotation of Human Protein-Coding Splice Variants Using Multiple Instance Learning.

Area of Science:

• Molecular Biology
• Bioinformatics
• Genomics

Background:

• Alternative splicing generates numerous protein-coding splice variants (PCSVs) from single genes, significantly expanding proteomic functional diversity.
• Current gene-level functional annotation algorithms lack isoform-level standards, limiting machine learning approaches.
• Abundant RNA-seq data enables detailed examination of gene function at the isoform level.

Purpose of the Study:

• To develop and validate a machine learning method for predicting the functions of protein-coding splice variants (PCSVs).
• To address the limitations of gene-level functional annotation by focusing on isoform-specific functions.
• To create a publicly accessible web resource for PCSV functional prediction.

Main Methods:

• Utilized a multiple instance learning (MIL) framework for PCSV function prediction.
• Integrated transcript-level expression data with gene-level functional associations from the Gene Ontology.
• Employed a support vector machine (SVM) with 5-fold cross-validation for model training and evaluation.

Main Results:

• The MIL-based approach demonstrated effective prediction of PCSV functions.
• Performance improved with genes possessing multiple PCSVs and with increased training data.
• Predictions were validated using literature evidence for specific genes (e.g., ADAM15, LMNA/C, DMXL2).

Conclusions:

• The developed method enhances functional annotation accuracy at the isoform level.
• The study highlights the importance of considering alternative splicing in functional genomics.
• The 'IsoFunc' web resource provides a valuable tool for exploring PCSV functions.